The Xiaomi AI Cube prototype combines the Xring O3, O100 and D100 accelerators and operates at a sustained 150W. It is built specifically for running AI models locally, and Xiaomi has shown it can handle a model with up to 120B parameters.
A standout feature is the O100 chip, which delivers 1.22 TB/s of near‑memory bandwidth thanks to its 3D‑stacking architecture. This bandwidth level highlights Xiaomi’s strong focus on optimizing inference directly on the device.
If Xiaomi can bring a compact, consumer‑ready version of this machine to market—capable of running models ranging from tens to hundreds of billions of parameters at home with reasonable performance and power costs—the local AI landscape could heat up very quickly.
Source Images
Conclusion
The Xiaomi AI Cube demonstrates that powerful, large‑scale AI inference can be packaged into a small form factor, and its commercial release could accelerate the growth of home‑based AI solutions.
